Efficient GPU implementation of parameter estimation of a statistical model for online advertisement optimization

Abstract

The optimization problem of estimating parameters using a maximum a-posterior (MAP) [3] approach on a non-linear statistical model with a large data set can be solved using an L-BFGS [10] algorithm. When dealing with an ever changing reality, the evaluation need to be fast to capture the immediacy of the observations. This thesis will present the implementation of the problem objective function and its gradient being used in the numerical iterative optimization algorithm. In order to speed up the process of parameter estimation, an implementation is presented which utilizes the massively parallel computation power of a graphics processing unit (GPU). The implementations are done for both the CPU and the GPU, using C++ and NVIDIA's programming platform CUDA. Compared to the sequential CPU implementation, the result of the parallel GPU version is a speed up of between 20 and 50 for the objective function and around 4 for the gradient.